shootthesound/comfyUI-Realtime-Lora

Train and block edit and save LoRAs directly inside ComfyUI for Z-image, Flux Klein, SDXL, Flux, WAN 2.2, SD 1.5

What it solves

This toolkit provides a unified interface within ComfyUI to train, analyze, and selectively load LoRAs (Low-Rank Adaptation) for a wide variety of image and video generation models. It eliminates the need to switch between different external training software and allows users to fine-tune how specific blocks of a model are influenced by a LoRA, providing much more granular control over the final output.

How it works

The project integrates three major training backends—sd-scripts, Musubi Tuner, and AI-Toolkit—directly into ComfyUI nodes. It offers a suite of tools for different stages of the model pipeline:

  • Training: Dedicated nodes for models like SDXL, SD 1.5, FLUX, Z-Image, Qwen Image, and Wan 2.2, allowing on-the-fly training from images.
  • Analysis & Selective Loading: "Analyzer" nodes detect the LoRA format (e.g., Standard, LoKR, LoHa) and identify which blocks have the most impact. "Selective Loader" nodes then allow users to toggle or adjust the strength of individual blocks or layers.
  • Strength Scheduling: V2 nodes allow LoRAs to be applied at varying strengths using keyframes and presets (e.g., fades, pulses).
  • Deep Debiasers & Inspectors: These nodes go beyond block-level control to scale individual sub-components (attention heads, MLPs, norms) of the DIT, VAE, and text encoders to fix biases or fine-tune behavior.

Who it’s for

Digital artists, AI researchers, and ComfyUI users who want to train custom subjects or styles and have precise, surgical control over how those LoRAs are applied to their base models.

Highlights

  • Multi-Backend Support: Unified access to sd-scripts, Musubi Tuner, and AI-Toolkit.
  • Broad Model Compatibility: Supports SDXL, SD 1.5, FLUX (including Klein 4B/9B), Z-Image, Qwen Image, and Wan 2.2.
  • Surgical Control: Block-level and sub-component-level (Deep Debiaser) editing for DITs, VAEs, and text encoders.
  • Dynamic Application: Strength scheduling with 40+ presets for evolving LoRA influence during generation.
  • Integrated Workflow: Train and apply LoRAs without leaving the ComfyUI interface.

Related

  • Project
    vrgamegirl19/comfyui-vrgamedevgirl

    A collection of ComfyUI custom nodes featuring an AI Video Builder for scene-by-scene production, video enhancement, and LoRA training for AI video and music video creation.

  • Project
    ussoewwin/ComfyUI-QwenImageLoraLoader

    A ComfyUI custom node suite for applying LoRAs and ControlNet to Nunchaku Qwen Image and Z-Image-Turbo models, featuring multi-LoRA stacking and Krea2 support.

  • Project
    willmiao/ComfyUI-Lora-Manager

    A comprehensive management tool for ComfyUI that streamlines the organization, downloading from CivitAI, and application of LoRA models and checkpoints.

  • Project
    Lightricks/ComfyUI-LTXVideo

    A collection of custom ComfyUI nodes and workflows that extend the LTX-2 video generation model with advanced control, HDR output, AI dubbing, and generative upscaling.

  • Project
    yolain/ComfyUI-Easy-Use

    ComfyUI‑Easy‑Use is a genuine open‑source extension for the ComfyUI node‑based UI, providing dozens of convenience nodes (loaders, prompt tools, LoRA/ControlNet helpers, noise injection, image utilities, logic/loop nodes, etc.) that streamline building Stable Diffusion and other diffusion model workflows. It supports many models (SD‑1/2, SDXL, SD‑3, Flux, Kolors, Stable Cascade, etc.), offers UI theme tweaks, and includes documentation, video tutorials, and a workflow collection. Installation is a simple git clone into ComfyUI’s `custom_nodes` folder followed by running `install.bat`.